UK edition
Founder, Trend Wire Group
Yash Malviya founded Trend Wire Group and edits Super Intelligence News. He started the title because most AI coverage rewards hype over substance, and he sets the rule the desk writes to: a development earns a story only when there is a primary source behind it, a paper, a model card or an official filing, and a plain reason it matters. He keeps the coverage on what the systems can actually do and what the evidence shows, not on what is trending that week.
Background: Founder of Trend Wire Group; sets editorial standards across its titles, verifies each claim against its primary source (papers, model cards, official announcements), and edits AI-assisted drafts to publication.
SWE-bench, SWE-bench Verified and LiveCodeBench are the scores labs quote to prove coding ability. They measure something real and narrow, and contamination, weak tests and saturation mean a high number says less about engineering reliability than it looks.
Red-teaming and safety evaluations increasingly decide which models ship. We look at the NIST framework, the UK and US safety institutes and METR to see what the tests measure, and why 'it passed our evals' says less than it sounds.
Britain's AI strategy mixes real Treasury spending with large private pledges under one banner. Separating the two shows a modest but genuine public commitment, a compute target restored after being cut, and a safety body that was renamed and funded but given no binding powers.
The headline is that AI will break the grid. The measured numbers are smaller and more interesting: data centres are about 1.5 percent of global electricity today, and the real fight is over how fast that doubles, and where.
Six AI growth zone announcements are listed by government, from Oxfordshire in April 2025 to Lanarkshire in February 2026. Most of the numbers are commitments, not built capacity.
Attorney General Rob Bonta served OpenAI with an investigative subpoena on 30 September 2026 over AI cybersecurity incidents. A federal probe and a 15 state coalition are already running.
Senators Hawley and Murphy introduced a bipartisan bill on 1 October 2026 that extends US computer crime law to the companies behind autonomous agents. The White House opposes it.
ASI is a hypothetical AI far smarter than the best human minds in practically every field. This week it became a US legal term. Nobody has agreed a test for it.
Anthropic's 30 September 2026 study finds near-universal technical exposure and almost no economic displacement today. The gap between the two numbers is the story.
Google DeepMind's new frontier model costs $2 per million input tokens and outputs up to 1 million tokens. It is limited to vetted defenders, and no third party has verified its scores.
Anthropic and the US government's CAISI both rate Zhipu's GLM-5.3 the strongest downloadable model for hacking tasks. The safeguards come off for about £3,300.
From 30 September 2026 the UK data regulator is run by a board, not one commissioner. Its powers are unchanged, but it is the body that must write the rules on AI and automated decisions.